🚨 NeurIPS Spotlight 🚨 #NeurIPS2025
How can we handle non-Markovian time-evolving offsets at test time in Offline Reinforcement Learning?
Tired of retraining and hyperparameter optimization when a new non-stationary pattern arises?
Introducing FORL: Forecasting in Non-stationary Offline RL
Towards real-world non-stationarity: We bridge the gap between offline RL and forecasting. 👇
📍San Diego Convention Center Exhibit Hall C,D,E #202
📅 11:00-14:00, Fri, Dec 5
📄 Forecasting in Offline Reinforcement Learning for
Non-stationary Environments https://t.co/z29G0bdDyC
🌐 https://t.co/pZKApiyRd7
Why is this important?
✨ LLMs: Semantic/contextual bias.
💊📈 Healthcare & Finance: Data may be withheld (privacy/regulations).
🦾 Industrial Robots: Sensors drift, requiring daily calibration.
FORL unifies:
1️⃣ Conditional Diffusion Models: Generate candidate states conditioned on action-effect (Δo, a) history.
2️⃣ Zero-shot Forecasting: Uses time-series foundation models.
3️⃣ Dimension-wise Closest Match: A lightweight approach to fuse information from both sources (no hyperparameters & no fallback mechanism)
How can an agent infer its location when the state is subject to large unknown offsets?
This action-effect (Δo, a) history is invariant to the offset. We train the diffusion model to generate plausible states conditioned on this history.
✅ No need to retrain when new patterns arise.
✅ Bypasses tedious hyperparameter optimization.
✅ Trained only on standard offline RL datasets.
We validated FORL on a novel benchmark integrating widely used time-series data and offline RL datasets on:
✅ No access to past offsets
✅ Inter/intra-episode drifts
✅ Policy-agnostic plug-and-play
✅ Offset magnitude-scaling
Suzan Ece Ada @suzaneceada, Georg Martius @GMartius, Emre Ugur @emreugur__, Erhan Oztop @ErhanOztop
🎥 https://t.co/WOlbEUyOHp
#NeurIPS #NeurIPS2025 #ReinforcementLearning #OfflineRL #DiffusionModels
@colorsLab_boun üyesi olan Dr. Suzan Ece Ada, doktorasını Emre Uğur @emreugur_ danışmanlığında ve Erhan Öztop eş-danışmanlığında yürütmüş olup doktora sonrası araştırmalarına ETH Zürich’te devam etmektedir. Mezunumuzu ve hocalarımızı çok tebrik ederiz 🎉
Dr. Ada’nın “Robust and Adaptive Deep Reinforcement Learning” başlıklı doktora tezi, Boğaziçi Üniversitesi Bilimsel Araştırma Projeleri Komisyonu tarafından üstün bilimsel niteliğe sahip olması nedeniyle Doktora Tez Ödülü’nü almaya layık görülmüştür.
I’m presenting FORL!
📍San Diego Convention Center Exhibit Hall C,D,E #202
📅 11:00-14:00, Fri, Dec 5
📄 Forecasting in Offline Reinforcement Learning for
Non-stationary Environments https://t.co/OOFDfBYW8I
🌐 https://t.co/iyFbM1uLbp…
🚨 NeurIPS Spotlight 🚨 #NeurIPS2025
How can we handle non-Markovian time-evolving offsets at test time in Offline Reinforcement Learning?
Tired of retraining and hyperparameter optimization when a new non-stationary pattern arises?
Introducing FORL: Forecasting in Non-stationary Offline RL
Towards real-world non-stationarity: We bridge the gap between offline RL and forecasting. 👇
📍San Diego Convention Center Exhibit Hall C,D,E #202
📅 11:00-14:00, Fri, Dec 5
📄 Forecasting in Offline Reinforcement Learning for
Non-stationary Environments https://t.co/z29G0bdDyC
🌐 https://t.co/pZKApiyRd7
Why is this important?
✨ LLMs: Semantic/contextual bias.
💊📈 Healthcare & Finance: Data may be withheld (privacy/regulations).
🦾 Industrial Robots: Sensors drift, requiring daily calibration.
FORL unifies:
1️⃣ Conditional Diffusion Models: Generate candidate states conditioned on action-effect (Δo, a) history.
2️⃣ Zero-shot Forecasting: Uses time-series foundation models.
3️⃣ Dimension-wise Closest Match: A lightweight approach to fuse information from both sources (no hyperparameters & no fallback mechanism)
How can an agent infer its location when the state is subject to large unknown offsets?
This action-effect (Δo, a) history is invariant to the offset. We train the diffusion model to generate plausible states conditioned on this history.
✅ No need to retrain when new patterns arise.
✅ Bypasses tedious hyperparameter optimization.
✅ Trained only on standard offline RL datasets.
We validated FORL on a novel benchmark integrating widely used time-series data and offline RL datasets on:
✅ No access to past offsets
✅ Inter/intra-episode drifts
✅ Policy-agnostic plug-and-play
✅ Offset magnitude-scaling
Suzan Ece Ada @suzaneceada, Georg Martius @GMartius, Emre Ugur @emreugur__, Erhan Oztop @ErhanOztop
🎥 https://t.co/WOlbEUyOHp
#NeurIPS #NeurIPS2025 #ReinforcementLearning #OfflineRL #DiffusionModels
Why does most of hierarchical RL stop at the first layer of hierarchy—i.e. skills? Because until now, it wasn't clear how to learn the state abstractions in a principled way! Our recent NeurIPS paper shows how to start with a set of skills and learn state abstractions. 🧵1/8
@emreugur__@colors_lab_boun@Bogazici_CmpE Sizinle ve Erhan hocamla çalışmak büyük bir ayrıcalıktı, çok teşekkür ederim hocam! Değerli jüri üyelerim, İnci Baytaş, Arzucan Özgür, Fatma Güney, Özgür Öğüz hocalarıma da çok teşekkür ederim 🙏🙏
@colors_lab_boun@Bogazici_CmpE öğrencim Suzan Ece Ada başarılı bir sunum ile doktorasını almaya hak kazandı. Robotica dergisi ile başladığı yayınlarını, 2 IEEE RA-L ve 1 Access makalesi ekledikten sonra NeurIPS Spotlight 🥳 ile en ileri noktaya taşıdı. Çok tebrikler Dr. Ada!
@svlevine was just presenting in the Exploration in AI @ #ICML2025 and promoted that exploration needs to be grounded, and that VLMs are a good source ;-) Check our paper below
👇
Zero-shot imitation from just a single sparse demonstration is hard. Goal-conditioned methods tend to “greedily" move from one state to the next and lose the big picture.
We're presenting an alternative approach on Tuesday at #ICML2025.
(1/3)
When multiple tasks need improvements, fine-tuning a generalist policy becomes tricky. How do we allocate a demonstration budget across a set of tasks of varied difficulty and familiarity?
We are presenting a possible solution at ICML on Wednesday!
(1/3)
✨Introducing SENSEI✨ We bring semantically meaningful exploration to model-based RL using VLMs.
With intrinsic rewards for novel yet useful behaviors, SENSEI showcases strong exploration in MiniHack, Pokémon Red & Robodesk.
Accepted at ICML 2025🎉
Joint work with @cgumbsch
🧵
All forms of intelligence co-emerged with a body, except AI
We're building a #future where AI evolves as your lifelike digital twin to assist your needs across health, sports, daily life, creativity, & beyond...
https://t.co/QL3o9YxZYz ➡️ Preview your first #HumanEmbodiedAI